Unsupervised Learning for Blind Source Separation


Abstract:

     We have previously applied Artificial Visual Systems (AVS) to image de-mixing, image 
compression combined with wavelet transforms, RGB-color vision, de-mix Super-Mother 
wavelet in terms of independent mother wavelets, as well as Continuous Speech Segmentation 
of phoneme-hyphenation. In this paper, after we have reviewed Independent Component 
Analysis (ICA) and Blind Source Separation (BSS) from the AVS viewpoint, we explore an 
intelligent pair of cameras in a time-multiplex fashion the possible condition to de-mix more 
than two scenes.  During the first look, the camera system has a plenty of time to have two
cameras pointed, triangulated, and focused perfectly on the girl_G standing in the shadow of
a tree_T hidden with birds_B. ?Suddenly the birds begin to sing. By design, the camera system 
has two independent pointing & focusing & memory capability to quickly response to the birds? 
song direction.  Therefore, only one of the dominate camera can point away from the girl_G and 
focus at the birds hidden in the tree at the second look. By design, the not-yet-in-focus and 
not-yet-correctly-pointed camera has replenished its blurred input with the previous de-mixed 
background image (T+B). Together, the second look can de-mix the bird_B from the Tree_T.  
We consider channel communication application that we can efficiently mix four images using 
matrices [A0] and [A1] to send through two channels.

Biography:

     Harold Szu is graduated in Physics from National Cheng Kung University, Taiwan, in 1963, 
has earned a Ph.D. degree in 1971 in Physics under the guidance of Professor George Uhlenbeck 
of the Rockefeller University, New York N.Y.  He is the head of Information Science Group in at 
Naval Surface Warfare Center (NSWC).  Dr. Szu has worked 15 years at Naval Research Lab.  He 
has been a visiting member of Institute of Advanced Studies, Princeton NJ from 1976-78.  He is 
a Founder, Past President, and current Editor-in-Chief of International Neural Network Society 
respectively in 1987, 1993-94.  During 1996-1998, he was Lamson Professor and responsible 
in the establishment of a NASA Earth Station at Lafayette Louisiana called NASA- Regional Data 
Validation Center. He has been published over three hundred papers, owned 8 patents, awarded 
with: the Fellow of SPIE in 1995, the Fellow of Optical Society of America in 1996, the Fellow 
of IEEE in 1997, and recipient of Gabor Prize of information science in 1997.  He has initiated 
and organized the Annual SPIE Orlando Wavelet Applications Conference since 1995.


[HOME]